Search

Peer-to-peer (P2P) lending is an emerging field in FinTech and is an alternative source of personal loans. However, P2P lending faces severe credit risk due to high information asymmetry and insufficient collateral. We develop a novel heterogeneous stacking ensemble (HSE) approach by using two real-world datasets to improve the loss given default (LGD) forecasting in the P2P lending domain. Some special data in P2P lending and macroeconomic variables are employed as supplementary data sources to further enhance the model performance. Our proposal is compared with several popular models, including parametric and non-parametric ones, in terms of predictive accuracy and capital requirement. Our finding reveals that special data in P2P lending (e.g., number of investors and loan description) and macroeconomic variables are powerful predictors of LGD in P2P lending. The proposed HSE model outperforms the benchmark models in most cases and significantly achieves optimal average ranks across all the evaluation metrics. The results remain robust under several validations.

On-road vehicle tracking is crucial for the development of automated vehicles and advanced driver assistance systems. Most existing tracking algorithms assume that the vehicles move independently without considering the interactions between surrounding vehicles. However, in order to keep a safe distance and avoid collisions, the motion behavior of vehicles should be affected by the surrounding vehicles. To address this limitation, this paper presents a novel vehicle tracking method that comprehensively considers the motion dependence interactions between vehicles. In this paper, the environment interaction models of lateral and longitudinal are constructed and incorporated into the estimation process that model the interaction behaviors in a 2-D road coordinate system. The problem of motion dependence estimation is handled by an interacting multiple model, which operates different dynamic models in parallel. Furthermore, to adapt the switching between the lateral maneuvering models, a Quasi-Bayesian recursion algorithm with adaptive transition probability is proposed. This adaptive switching strategy improves the accuracy of lateral motion behavior estimation and thus improves the tracking performance. The proposed algorithm is evaluated by simulation, and its performance is quantified by the posterior Cramer-Rao lower bound. The simulation results show that the presented algorithm provides higher accuracy and reliability after considering the environment interaction.

Since single sensor and high-density point cloud data processing have certain direct processing limitations in urban traffic scenarios, this paper proposes a 3D instance segmentation and object detection framework for urban transportation scenes based on the fusion of Lidar remote sensing technology and optical image sensing technology. Firstly, multi-source and multi-mode data pre-fusion and alignment of Lidar and camera sensor data are effectively carried out, and then a unique and innovative network of stereo regional proposal selective search-driven DAGNN is constructed. Finally, using the multi-dimensional information interaction, three-dimensional point clouds with multi-features and unique concave-convex geometric characteristics are instance over-segmented and clustered by the hypervoxel storage in the remarkable octree and growing voxels. Finally, the positioning and semantic information of significant 3D object detection in this paper are visualized by multi-dimensional mapping of the boundary box. The experimental results validate the effectiveness of the proposed framework with excellent feedback for small objects, object stacking, and object occlusion. It can be a remediable or alternative plan to a single sensor and provide an essential theoretical and application basis for remote sensing, autonomous driving, environment modeling, autonomous navigation, and path planning under the V2X intelligent network space– ground integration in the future.

Large-scale knowledge graphs (KGs) play a critical role in question answering over KGs (KGs-QA). Despite of large scale, KGs suffer from incompleteness which results in the low-precision of KGs-QA. Since existing researches of relation prediction process each triple independently, the hidden relations which are inherently present in the neighborhood can not be captured. Complementarily, to capture both entity and relation features in a given entity's neighborhood, an entity importance estimation network of attention-based graph embedding is proposed, which consists of the attention-based graph embedding module and the entity importance estimation module. The new embedding of an entity from its n-hop neighbor is learned by an attention-based graph embedding module, and then the learned embedding is integrated into the entity importance estimation module to find entities of high importance in n-hop neighbors of the central entity. Multi-hop relations are encapsulated and an auxiliary edge of n-hop neighbors is introduced, which realizes the relation prediction task. In addition, we are the first to realize KGs-QA while realizing relation prediction, which alleviates the phenomenon of missing relations and the low-precision problem of KGs-QA. As expected, KGs-QA's F1 increases significantly when relation prediction is invoked. In 80% missing relation, we observe a F1 increase of 21, 17, 18 and 16 points over the no-relation prediction configuration for SQ, WQ, GQ and QALD-5, respectively.

Continuous point cloud stitching provides accurate and rapid environmental information for mapping, which plays an essential role in autonomous vehicles. Many methods related to autonomous vehicles are either inaccurate or time-consuming when they reconstruct large scenes. In this work, a feature matching score-based stitching method to reconstruct environments for autonomous vehicles is proposed. The purpose of the method is to address the problems of stitching quality and time consumption by using the feature match- completeness scores. The feature matching score is obtained by calculating the matching rate within an adjacent image pairs that produces a frame of point cloud. The score maintains the correct points, so fewer points require to be stitched. Then, the corresponding criterion is given to evaluate the performance of stitching candidate point clouds. Finally, the points are fit into planes to form a complete and comprehensive scene. The proposed stitching algorithm was verified in several experiments to demonstrate the efficiency and the accuracy of the overall system.

Peer‐to‐peer (P2P) lending is facing severe information asymmetry problems and depends highly on the internal credit scoring system. This paper provides a novel credit scoring model, which forecasts the probability of default for each applicant and guides the lenders' decision‐making in P2P lending. The proposal is expected to improve the existing credit scoring models in P2P lending from two aspects, namely the classifier and the usage of narrative data. We utilize an advanced gradient boosting decision tree technique (i.e., CatBoost) to predict default loans. Moreover, a soft information extraction technique based on keyword clustering is developed to compensate for the insufficient hard credit data. Validated on three real‐world datasets, the experimental results demonstrate that variables extracted from narrative data are powerful features, and the utilization of narrative data significantly improves the predictability relative to solely using hard information. The results of sensitivity analysis reveal that CatBoost outperforms the industry benchmark under different cluster numbers of extracted soft information; meanwhile a small number of clusters (e.g., three) is preferred for consideration of model performance, computational cost, and comprehensibility. We finally facilitate a discussion on practical implication and explanatory considerations.

Vehicle tracking based on fusion of radar and camera is one of most important approaches for intelligent vehicle, and data association playes key roles in multi-target tracking. To improve the performance of data association to increase the tracking accuracy, a new vehicle tracking method for radar and monocular camera is proposed. By the method, the tracking gate in polar coordinates is designed to acquire valid measurement firstly. Considering that the poor distance precision of monocular camera would introduce incorrect association between measurements and target tracks in Cartesian coordinates, only angle of target is utilized for data association in polar coordinates to obtain proper match. Then the filtering and tracking accuracy is increased by more correct data association. The results of experiment show that the proposed vehicle tracking method can improve the vehicle tracking accuracy, and reduce the computational costs.

The urban black and odorous water bodies affect the daily life of urban residents, destroy the aquatic ecosystem and damage the image of the city. This research used the evaluation method of black and odorous water body proposed by the Ministry of Housing and Urban-Rural Development of the People’s Republic of China (MOHURD) to evaluate the water quality of 50 key creeks (53 sections) in Guangzhou in 2016 and 2017 and analyzed the spatial distribution of black and odorous water bodies and their changes over time. The results showed that the water quality of Conghua District, Panyu District, Nansha District, and Yuexiu District was relatively good, while the water quality of Baiyun District, Haizhu District, Huadu District, Huangpu District, Liwan District, Tianhe District, and Zengcheng District was relatively poor. Compared with 2016, the water quality of creeks in Huangpu District and Baiyun District deteriorated in 2017, while the water quality of creeks in Tianhe District, Liwan District, and Huadu District improved. The water quality of creeks in other districts did not change significantly. In 2017, ammonia nitrogen (NH3-N), total phosphorus (TP), and chemical oxygen demand (COD) content decreased, and dissolved oxygen (DO) content increased compared to 2016. In 2017, the number of black-odor water bodies was greatly reduced, and the water quality of creeks was generally improved. The results can provide a reference for the treatment of the black and odorous creeks in Guangzhou.

This paper proposes a new car-following (CF) model incorporating the effects of lateral gap and roadside device communication to capture the characteristics of electric vehicle (EV) traffic stream in transportation-cyber-physical systems. Stability of the proposed CF model is analyzed using the perturbation method. Furthermore, the energy consumption of the EV traffic stream is investigated based on the drive cycles produced by the proposed model. Numerical experiments analyze three scenarios: start, stop, and evolution processes for the scenarios of no lateral gap, lateral gap, and lateral gap with roadside device, respectively. Results demonstrate that: 1) the nonlane-discipline-based model is more responsive than the lane-discipline-based model; 2) the nonlane-discipline-based model for the EV traffic stream consumes more energy in the acceleration phase and recuperates more energy in the deceleration phase compared with the lane-discipline-based model; and 3) the nonlane-discipline-based model with roadside device communication for EV traffic stream consumes more energy in the acceleration phase and recuperates more energy in the deceleration phase than the model without roadside devices.

Booming online consumer lending encounters high credit risk and thus needs well-designed credit scoring models. As a supplementary data source, multilevel macroeconomic variables (MVs), which means mixed regions, data frequency, and lags of MVs, are combined with application data for credit scoring. Moreover, we propose a Bayesian MV selection and lag optimization method to handle the highly correlated MVs and capture flexible lag effects. A structural model for consumer default behavior is proposed to analyze the potential role of multilevel MVs in theoretically predicting default loans. Validated on a real-world dataset from an anonymous Chinese online consumer lending platform, the empirical results confirm that multilevel MVs are significant determinants of consumer credit risk. The comparison of classification performance further implies the superiority of incorporating multilevel MVs and the proposed MV selection and lag optimization method relative to benchmarks. Gradient boosting decision tree-based approaches provide significantly better performance than industry benchmarks while maintaining interpretability via Shapley additive explanations values.

As self-driving vehicles have grown in sophistication and ability, they have been deployed on the road in regular private vehicles. This paper presents a new semi-supervised multi-granularity convolutional neural networks (SSMGCNNs) framework for semantic analysis in designing human-vehicle interaction, which is further composed of two parts: a two-view-embedding (TVE) module, and the multi-granularity convolutional neural networks (MGCNNs) module. SSMGCNNs learn embeddings of text regions from the unlabeled user instruction dataset and then integrate the learned tv-embedding into MGCNNs, so that the learned tv-embedded regions are used as an additional input to MGCNNs convolution layer to solve the problem of data annotation. MGCNNs can fully extract feature information hidden in text through multiple convolution kernels of different sizes in the same convolution layer. The experimental results demonstrate the effectiveness of our approach in comparison with baseline systems. ...

Driving environment information perception of intelligent vehicle is a major foundation for active and safe automatic driving. Perception of 3D object detection is an important basis for intelligent behavior decision-making such as obstacle recognition, collision avoidance, path planning and etc. In this paper, we present a multi-layer structured object detection framework under parallel pipeline. Our method does not rely solely on pure 3D scheme, but leverages region-proposal 2D learning classification detector and volumatic point cloud over-segmentation under 3D physical and geometric attributes for object localization. We innovative reduce the search space of 3D detection, implement multi-mode data fusion and multi-dimensional information interaction. Even scenes with different scales and different granularity knowledge can get corresponding 3D object detection and semantic information perception. Meanwhile, input with sparse data or stacked occlusion will also output certain detection feedback. Experiments show that the work of this paper can improve the accuracy and feasibility of segmentation of 3D instances in complex scenes of intelligent vehicles, and output cognitive results with location information and semantic perception information.

Aiming to achieving safe and efficient energy utilization for electric vehicles, research into the monitoring of lithium-ion batteries (LIBs) has become increasingly important. However, various estimation strategies are proposed at the cost of the higher design complexity and the poorer model performance, which are hard to be implemented. Complementarily, in this paper, we propose an Deep Neural Networks (DNNs)-based State of Charge (SOC) observer design for LIBs to ensure safe and reliable battery operations, which avoiding overcharging or over-discharging of the battery. More specifically, a Recursive Neural Networks (RNNs)-based feature extraction model is proposed to obtain sufficient feature information. Then, the well-trained feature vector is integrated into Convolutional Neural Networks (CNNs) to predict the LIBs SOC. In other words, the output of the RNNs are used as the input of the CNNs, which this practice can improve the model performance obviously. Furthermore, the extensive real-world experiments demonstrate that Neural Network-based SOC prediction model can provide faster convergence speed and higher precision in contrast to the optimal method to achieve SOC estimation over regular model.

Here, real-time 6-DOF monocular visual SLAM, which can be mapping with texture, is presented. The proposed method aims to compute 6-DOF camera pose and 3D landmarks position using monocular camera as the only sensor, and utilize texture to render the map of visualization quickly. By exploiting additional structural information such as camera height from the ground, the epipolar geometry is transformed into the perspective-n-point problem, restoring the indeterminate scale. In addition, this structural information is used to segment the point cloud of ground. This is of great help to clustering obstacles and constructing the spatial segmentation relationship of point cloud. The real-time performance of system is demonstrated using only GPU to render map. The effectiveness of the proposed method is demonstrated on various sequences including the KITTI dataset and outdoor image sequences captured on experimental vehicle.

This study proposes a sliding mode controller for vehicular traffic flow based on a car-following model to enhance the smoothness and stability of traffic flow evolution. In particular, the full velocity difference (FVD) model is used to capture the characteristics of vehicular traffic flow. The proposed sliding mode controller is designed in terms of the error between the desired space headway and the actual space headway. The stability of the controller is guaranteed using the Lyapunov technique. Numerical experiments are used to compare the performance of sliding mode control (SMC) with that of feedback control. The results illustrate the effectiveness of the proposed SMC method in terms of the distribution smoothness and stability of the space headway, velocity, and acceleration profiles. They further illustrate that the SMC strategy is superior to that of the feedback control strategy, while enabling computational efficiency that can aid in practical applications.

This study proposes a new lattice hydrodynamic model considering the effects of bilateral gaps on a road without lane discipline. In particular, a lattice hydrodynamic model is proposed to capture the impacts from the lateral gaps of the right-side and left-side sites of the considered lattice sites. Linear stability analysis of the proposed model is performed using the perturbation method to obtain the stability condition. Nonlinear analysis of the proposed model is performed using the reductive perturbation method to derive the modified Korteweg–de Vries (mKdV) equation to characterize the density wave propagation. Results from numerical experiments illustrate that the smoothness and stability of the proposed model are improved compared with the model that considers the effect of unilateral gap. Also, the proposed model is able to more quickly dissipate the effect of perturbation occurring in the vehicular traffic flow.

This study proposes a new car-following model considering the effect of electronic throttle opening angle to capture the characteristics of connected autonomous vehicular traffic flow. The proposed model incorporates the opening angle of electronic throttle into the full velocity difference (FVD) model by assuming that the information of electronic throttle dynamics is shared by vehicles through vehicle-to-vehicle communications. The stability condition of the proposed car-following model is obtained using the perturbation method. Numerical experiments are constructed on three scenarios, start, stop, and evolution processes, to analyze the vehicular traffic flow characteristics of the proposed model. The results of numerical experiments illustrate that the proposed car-following model has a larger stable region compared with FVD model. Also, it demonstrates that the proposed car-following model can better present the characteristics of connected and autonomous vehicular flow in terms of the smoothness and stability with respect to the space headway, position, velocity, and acceleration/deceleration profiles.

This study proposes a feedback-based platoon control protocol for connected autonomous vehicles (CAVs) under different network topologies of initial states. In particularly, algebraic graph theory is used to describe the network topology. Then, the leader–follower approach is used to model the interactions between CAVs. In addition, feedback-based protocol is designed to control the platoon considering the longitudinal and lateral gaps simultaneously as well as different network topologies. The stability and consensus of the vehicular platoon is analyzed using the Lyapunov technique. Effects of different network topologies of initial states on convergence time and robustness of platoon control are investigated. Results from numerical experiments demonstrate the effectiveness of the proposed protocol with respect to the position and velocity consensus in terms of the convergence time and robustness. Also, the findings of this study illustrate the convergence time of the control protocol is associated with the initial states, while the robustness is not affected by the initial states significantly.

Considering the chaotic characteristics of traffic flow, this study proposes a Bayesian theory-based multiple measures chaotic time series prediction algorithm. In particular, a time series of three traffic measures, i.e., speed, occupancy, and flow, obtained from different sources is used to reconstruct the phase space using the phase space reconstruction theory. Then, data from the multiple sources are combined using Bayesian estimation theory to identify the chaotic characteristics of traffic flow. In addition, a radial basis function (RBF) neural network is designed to predict the traffic flow. Compared to the consideration of a single source, results from numerical experiments demonstrate the improved effectiveness of the proposed multi-measure method in terms of accuracy and timeliness for the short-term traffic flow prediction.

This study proposes a control strategy of vehicular traffic based on the lattice hydrodynamic model by taking the difference of density change rate as a feedback signal. In particular, a delay feedback control scheme is designed based on Nagatani's lattice model considering the effects of density change rate difference under the open boundary condition. Stability analysis of the lattice hydrodynamic model with the control signal is performed using the small gains theorem to obtain the stability condition. Results from numerical experiments including the single perturbation and multiple perturbations illustrate that the proposed control strategy can suppress traffic jam effectively and make the vehicular traffic flow to move into a comparatively homogeneous state.

Environment perception is crucial for the development of autonomous driving and advanced driver assistance systems. The cooperative perception using the infrastructure sensors can significantly expand the field of view of on-board sensors and improve the accuracy of target tracking. In this paper, we propose a hybrid vehicular perception system that incorporates both received feature-level information from infrastructure sensors and track-level data from the multi-access edge computing server (MEC-Server). An infrastructure-enhanced multiple-model probability hypothesis density is proposed to handle the feature-level data from heterogeneous infrastructure sensors. The problem of kinematic state estimation is improved with the prior information of the road environment. Furthermore, a generic communication interface between the infrastructure sensor and MEC-Server is designed, which allows the object data to have the same notion of locality through the use of a generic object state model. Simulation results show that the presented algorithm provides higher accuracy and reliability after considering the prior information of the road environment.

Ensemble models have been extensively applied to credit scoring. However, advanced tree-based classifiers have been seldom utilized as components of ensemble models. Moreover, few studies have considered dynamic ensemble selection. To fill the research gap, this paper aims to develop a novel tree-based overfitting-cautious heterogeneous ensemble model (i.e., OCHE) for credit scoring which departs from existing literature on base models and ensemble selection strategy. Regarding base models, tree-based techniques are employed to acquire a balance between predictive accuracy and computational cost. In terms of ensemble selection, the proposed method can assign weights to base models dynamically according to the overfitting measure. Validated on five public datasets, the proposed approach is compared with several popular benchmark models and selection strategies on predictive accuracy and computational cost measures. For predictive accuracy, the proposed approach outperforms the benchmark models significantly in most cases based on the non-parametric significance test. It also performs marginally better than several state-of-the-art studies. Our proposal remains robust in several scenarios. In terms of computational cost, the proposed method provides acceptable performance and benefits from GPU acceleration considerably.

Based on the different sensitivities as drivers to response for the space headway or the velocity difference between the following vehicle and the leading vehicle, this study proposes a new car-following (CF) model that considers two different time delays in sensing space headway and velocity. The stability analysis of the proposed CF model is performed using the perturbation method and the root locus method to obtain the stability condition. Simulation-based numerical experiments are conducted. Results show that the stable region of the proposed model varies with the difference between two different time delays.

The environment perception plays a significantly role in intelligent vehicles and the advanced driver assistance system (ADAS), which enhances the driving safety and convenience. Target tracking is one of the key technologies of environment perception. The on-board sensors such as cameras and radar are commonly used for target tracking while they have limitations in terms of detection range and angle of view. One way to overcome the perception limitations of on-board ranging sensors by incorporating the vehicle-to-vehicle (V2V) communication. This paper proposes a vehicle tracking system which fuse the radar and V2V information to improve the target tracking accuracy. The proposed system integrates the radar, GPS and DSRC communication equipment. The GPS and radar are utilized to obtain its own position information and the position information of nearby vehicles. The proposed system also resolves the problem of data association in multiple target tracking measurements by other connected vehicles' identity information. With the association measurements, a Kalman filter is used to improve the accuracy of target tracking. An assessment of tracking system in real road environment shows that the proposed fusion approach for target tracking can reduce the data association error and improve the vehicle target tracking accuracy.

The purpose of environment perception for autonomous vehicle is to get the entire information around the ego-vehicle and understand the environment where the autonomous vehicle runs. In this paper, we attempt to rebuild a 3D environment for the autonomous vehicle using the vehicle mounted cameras. First, we generate the continuous 3D point clouds with the vehicle moving. Then we stitch the continuous point clouds by using the image feature matching constraint and scores. At last, we obtained an accuracy and efficiency point cloud of the environment. After the numerical experiments, the proposed stitching algorithm has been verified that the MSE is lower than the average of other algorithms. And the number of points decreases than the conventional algorithm.

3D Object Detection is important to avoid collision and path planning in field of autonomous vehicle. In this paper, we present a combined clustering and image mapping-based algorithm to segment 3D point cloud. It not only provides a dependable initial value as the seeds to cluster the class of objects, but also avoid the pre-trained classifier to detect the objects. We get an accurate 3D object detection result using our proposed algorithm. The proposed algorithm can reduce the computation complexity at the step of determining bounding area in 2D image and produce the initial center of cluster of each object at the step of segmentation in 3D point cloud. The experiment states that the proposed algorithm can improve the accuracy and feasibility of object detection.

The environment perception of autonomous vehicle is mainly aimed at obtaining the holonomic information around the ego-vehicle to understand the driving environment. In this paper, we attempt to use only vehicle-mounted cameras to reconstruct the 3D environment for the autonomous vehicle perception. Firstly, we acquire the continuous frames in vehicle motion to reconstruct continuous 3D point clouds. Secondly, the continuous point clouds can be stitched by the proposed matching constraint and scores of image features. Lastly, we get high-accuracy and high-efficiency dense point cloud of the environment. Experimental results on the benchmark dataset demonstrate the effectiveness and robustness of the proposed stitching algorithm. Compared with other algorithms and its variants, the MSE of the proposed method is lower than the average and the number of redundant points in overlap regions is reduced.